Temporal structure of the BN Modulation-filtered Spectrogram features (msg1a&b)

1999feb19 Dan Ellis <dpwe@icsi.berkeley.edu>

Introduction

In the course of preparing the slides for the DARPA Broadcast News workshop, I made some spectrogram-style plots based on the MSG features along with the standard PLP features for comparison.  I don't think I'd ever before looked at visualizations of the particular instance of MSG that we have settled on in the BN work, and it wasn't quite as I had expected:


The RASTA features (these are the 16ms-adjusted) are clearly blurred-out versions of the PLP, and the msg1b features do indeed look even more blurred in addition to more exotic phase effects, but the msg1a features are more than just blurred - they're ringing.  I thought maybe it was a bug, so I checked the modulation-domain filter response (a 17-point FIR filter), but it was perfect - flat from 0-16 Hz,
then ripples peaking at -40dB from about 20 Hz up.  It does, however, have ringing in the time domain.  Plotting just the 7th frequency channel for each of the above shows it very clearly:

And here are the frequency and impulse responses of the two filters, just to confirm:

Interesting, eh?  We've tried several different kinds of msg feature with BN, and these ones have always come out on top.



Back to Random Research Notes - DAn's Research - DAn's homepage